Regulations and compliance · Practical guide

Human review of AI output in GxP workflows

Start by specifying what the person will check. Do they need to verify completeness, technical accuracy, the cited source or a particular decision? Without a defined question, review can become little more than reading. A fluent answer can inspire more confidence than its contents warrant.

Table of contents

At a glance

When does a human reviewer provide an effective control over AI output?

A reviewer needs expertise, time and reference information to detect errors, correct them and prevent unsuitable output from being used. Test that the reviewer can exercise that authority in the actual workflow. In the 2025 Annex 22 consultation draft, human review does not create an exception to the proposed scope for critical applications. For non-critical uses, the draft places responsibility for output suitability with qualified personnel.

Scope and regulatory status

Use the following questions to assess how human review works in practice. References to Annex 22 concern the 2025 consultation draft. Assess the applicable framework and the intended use separately.

The EMA report published on 2 October 2026 records discussion of adding further model types. The revision is ongoing. A human signature alone does not establish either that a control works or that the proposed use is permissible.

Sources: [1] [2] [4]

Can the reviewer reject an output and explain why?

Give the reviewer independent reference information. Asking the same AI whether its first answer was right supplies no independent evidence. The reviewer needs approved originals and clear process criteria to check the output.

Test the control with known errors at a realistic workload. A reviewer asked to approve a hundred long outputs under time pressure may miss errors. Limit permitted use to the level of review people can achieve in practice.

Sources: [1] [2] [3]

Describe the control as a specific human responsibility

Describe what the reviewer will check, what evidence they can compare it against and what they are authorised to decide. Without those conditions, another signature provides little evidence that the control works.

Under section 3.3 of the 2025 Annex 22 consultation draft, where human decisions justify reduced model testing, the intended use should describe the operator’s responsibility. Training and consistent performance are then monitored as they would be for a manual process. Section 10.5 proposes documenting ongoing review, which may cover every output depending on criticality. These proposals do not expand the draft’s scope for critical applications. The exercises below are our own examples.

Distinguish the reviewer’s responsibility from technical operation of the model. A reviewer can assess an output without managing infrastructure or training data but needs to know the relevant limits. The model operator can explain a metric. Responsibility for the quality decision still sits with the designated process role.

Sources: [1] [2] [3]

Further reading:

Build a matrix of errors, reference information and expected responses

Derive error cases from the task. A summary can omit a restriction, a classifier can mark a critical case as normal and a numerical output can carry the wrong unit. For each case, specify how the reviewer can detect the error and what action should follow.

Test workload and presentation as well as access. A complete source can be impossible to use under time pressure. A side-by-side comparison can help if it retains the relevant information. Reproduce actual working conditions in the exercise.

Worked review matrix for detectable errors and their consequences
Error typeReference information neededExpected action
Incorrect fact Appropriate original source Correct or reject the output
Omitted exception Complete relevant context Identify the missing qualification
Incorrect unit The applicable rule and original input data Prevent further use until the value is checked
Uncertain classification Defined criteria and escalation route Escalate the borderline case to the appropriate person
Unsupported source Access to the actual source record Do not accept assertion as confirmed.

Sources: [1] [2] [3]

Testing known errors under a realistic workload

Give the reviewer ten short outputs containing two known factual errors and one omitted qualification. The exercise tests whether the available information and authority suit the task. Its result does not establish a universal success rate.

Set the expected findings and actions in advance. Give the reviewer the sources and time available in normal operation. Record what they detect, what they miss and which questions they need to ask. A debrief supports learning and does not change the observed result.

Repeat relevant cases after improving the display, criteria or training. Also test whether a correct routine output is rejected unnecessarily. The purpose of the control is to support the intended decisions. Determine the scope of the exercise and the evaluation from the task and its criticality.

Test cases for evaluating human error detection
Exercise caseCondition introducedExpected control
Normal case Technically correct output Accept with a reason within the permitted scope
Incorrect fact Plausible but incorrect statement Detect the error by checking the original
Omission Critical exception is missing Check the full context
Uncertainty A case that cannot be judged unambiguously Follow the escalation route when the evidence is inconclusive
Workload Realistic workload of AI outputs The reviewer can still complete the required checks

Sources: [1] [2] [3]

Give reviewers independent sources and authority to reject

A second answer from the same model can repeat an error and sound convincing. Give reviewers independent, approved sources and clear criteria. Specify which source takes precedence if documents conflict.

The reviewer must be able to reject an output or stop its use. If clicking “Confirm” is the only practical route, the control may fail. Define correction, escalation and handling of unresolved questions. Check the time required and the availability of cover.

High past model performance does not make an individual output correct. Equally, general mistrust is a poor reason to reject a result. Base the decision on specific criteria and relevant evidence.

Sources: [1] [2] [3]

Distinguish evidence of review from evidence of model performance

A signature documents that a review took place. To assess its quality, examine observed behaviour, cases with known errors and subsequent findings. Different critical error types can have different consequences. Avoid an unsupported universal percentage for acceptance.

Having a human in the process does not remove the need to assess the model. The draft discusses oversight in the context of reduced model testing and the criticality of the application. Demonstrate how well the control works in the specific system.

Record the scope, model identification, input and decision to the extent the process needs. A daily sign-off that cannot be linked to the outputs reviewed may offer little evidence. Keep records traceable and avoid unnecessary duplication.

Sources: [1] [2] [3]

Detect changes and gradual expansion of intended use

A new model, data source or task can change what the reviewer must detect. Reassess the available references and the reviewer’s expertise. Someone qualified to review wording may need different expertise to judge an analytical classification.

Observe how the tool is actually used. Staff might initially use text as a drafting aid and later copy it directly into a record. That changes the risk even if the software stays the same. Define permitted tasks and a route for assessing new use cases that staff can follow in routine work.

Keep reference cases for refresher training and change assessment. Include significant error types and update the cases as you learn more. The review should be a defined task that can detect errors, rather than a signature after every output.

Sources: [1] [2] [3]

Worksheet for testing human review under routine conditions

Define the reviewer’s task, available source, time and decision authority. In this example they compare a summary with the original report. The test includes accurate statements, one wrong number and an omitted qualification. Prepare the expected findings independently.

Observe how the reviewer reaches the decision. Can they find the relevant passage in the original? Can they recognise uncertainty and stop use? A control that looks sound in a procedure may fail when the display is unsuitable or the reviewer has too little time.

Apply the results only to the task tested. Success in reviewing text does not establish competence to assess a new analytical classification. Review changes in use and retain suitable reference cases. “A human checks everything” is too imprecise to describe an effective control.

Worked human-review worksheet
Control fieldExample entryWhat to observe
Task Compare the summary with the original report. Does the reviewer understand what to check?
Reference Independent, approved source. Is there evidence independent of the model’s own answer?
Error Incorrect number and missing exception. Are relevant differences identified?
Authority Reject or escalate the output. Can inappropriate use be prevented?
Workload A representative batch of outputs under routine conditions. Can the reviewer still perform the check?
Improvement Adjust the criteria, the way information is displayed or the training. Is the effect of the change tested again?

Sources: [1] [2] [3]

Next steps

  • Identify the claim to check and the consequences of error.
  • Clarify the reviewer's expertise, sources, and time requirements.
  • Test the control with known errors.
  • Define how reviewers reject or correct an output and record their decision.
Further reading:

Project experience and collaboration

Test human review with errors the reviewer can detect

With your process owners, we prepare sample outputs containing both plausible content and deliberately introduced errors. The team reviews them using the information and time available in practice. If reviewers reliably identify relevant errors, we can define the control more precisely and document its performance. Otherwise the task, information used or permitted AI use needs to be narrowed. The exercise replaces neither the operational decision on use nor the review of applicable requirements. Human review does not extend the scope described in the draft Annex 22 for critical applications.

Frequently asked questions about Human review of AI outputs

Is a second AI an independent technical reviewer?

A second model can repeat the same error and sound equally convincing. It can help find information, but the reviewer needs an independent original source for important facts and decisions. Verify that the reviewer can access and assess the authoritative information.

Sources: [1] [2]

Must every AI-generated text be completely rewritten?

No. The reviewer needs to detect errors that matter. Set the scope of review according to the content, intended use and risk.

Sources: [1] [2]

Does human-in-the-loop make an AI deployment automatically permissible?

No. Human oversight does not replace assessment of the intended use or the applicable requirements. For critical GMP applications, the 2025 draft Annex 22 excludes dynamic and probabilistic models, generative AI and large language models. Adding human approval does not change the scope of the draft. For an otherwise suitable use, demonstrate whether the reviewer can actually detect relevant errors and prevent the output from being used inappropriately.

Sources: [1] [2]

How can we test whether human review works?

Use routine cases and known errors under realistic conditions. Define the expected detection and action before the exercise. Record the decision, follow-up questions and missed cases. Improve the criteria, presentation or training where needed. The method depends on the task and its risks.

Can additional training fix every weakness in human review?

No. Limited time, inaccessible sources and lack of authority to reject can prevent effective review. Assess the actual task and improve the process or display where appropriate. Training can address a gap in knowledge or practice. It cannot give the reviewer access or authority they do not have.

Sources and review date

The model scope comes from the 2025 consultation draft. The EMA workshop report published on 2 October 2026 documents the subsequent discussion. The regulatory status was checked on 17 November 2026. We developed the assessment questions and examples.

  1. European Commission: EU GMP Annex 11, Computerised Systems (2011)EudraLex listed the January 2011 text as current on 14 October 2026. It addresses applications, risks, operation and the data lifecycle. The proposed tests are our own practical interpretation.
  2. European Commission: Annex 22 Artificial Intelligence – consultation draft 20252025 consultation draft. Its scope limits the model types proposed for critical GMP applications. Section 3.3 addresses operator responsibility when model testing is reduced. Section 10.5 addresses documented ongoing review. Human involvement does not expand the draft’s scope.
  3. European Commission: Consultation on Chapter 4, Annex 11 and Annex 22 (2025)Consultation ran from 7 July to 7 October 2025 and is closed. Publication for consultation does not establish entry into force.
  4. EMA · Workshop report on the development of Annex 22EMA/156789/2026, published on 2 October 2026. See the introduction on p. 3 and the next steps on p. 14.

Sources checked: .

Daniel Herrmann, Daniel Herrmann Consulting

Daniel Herrmann Consulting

Defining a human review process you can demonstrate

We help you assess the use case, assign responsibilities and capture evidence that the control works in operation. The appropriate support depends on your system, your team and the decision you need to make.

  • AI system validation →

    Bring your AI system into GxP use with traceable evidence and keep it under control in operation. Once the system and intended GxP use are defined, we validate according to risk, addressing data, supplier, tests, release and change control.

Send a project enquiry Customer testimonials →